CompVis / latent-diffusion

High-Resolution Image Synthesis with Latent Diffusion Models
MIT License
11.99k stars 1.54k forks source link

What is 'ADM' mean in this paper? #294

Open choidaedae opened 1 year ago

choidaedae commented 1 year ago

I am an undergrad student in CS, and I have been interested in Generative model recently. I apologize for my ignorance.

I have a question while reading your paper that contributed greatly to the diffusion model. In an experiment comparing FIDs of generated images with other diffusion models, or generative models such as GAN, I thought LDM-1 was ADM (A Diffusion model without classifier guidance in paper: "Diffusion model beat GANs ~~". However, in this experiment, LDM-1 was compared with ADM, and ADM performed better. What is the 'ADM' used in this experiment? Can you give me the name or link of the paper that was presented in which paper?

moonbow721 commented 1 year ago

"Diffusion models beat GANs on image synthesis", I think.

choidaedae commented 1 year ago

As I mentioned above, at first, I also thought that it is a model introduced in that paper, but In experiment in this paper, LDM-1 is being compared with that ADM. Why aren't they same? Which models were used in LDM-1 and ADM?

ZiQi-Jiang commented 1 year ago

guided diffusion, from Diffusion Models beat GANs on image synthesis

SanjeevaRDodlapati commented 11 months ago

ADM - ablated diffusion model

Harr7y commented 7 months ago

LDM: Latent diffusion model -> diffusion occurs at latent representation. ADM: Ablated diffusion model -> diffusion occurs at image level.